github/OpenHands
1
0
Fork 0
mirror of https://github.com/OpenHands/OpenHands.git synced 2025-12-26 13:52:43 +08:00
Code Issues Packages Projects Releases Wiki Activity
OpenHands/opendevin/server
History
Robert Brennan 1e95fa435d
Microagents and Delegation (#1238) 
* basic microagent structure

* start on jinja

* add instructions parser

* add action instructions

* add history instructions

* fix a few issues

* fix a few issues

* fix issues

* fix agent encoding

* fix up anon class

* prompt to fix errors

* less debug info when errors happen

* add another traceback

* add output to finish

* fix math prompt

* fix pg prompt

* fix up json prompt

* fix math prompt

* fix math prompt

* fix repo prompt

* fix up repo explorer

* update lock

* revert changes to agent_controller

* refactor microagent registration a bit

* create delegate action

* delegation working

* add finish action to manager

* fix tests

* rename microagents registry

* rename fn

* logspam

* add metadata to manager agent

* fix message

* move repo_explorer

* add delegator agent

* rename agent_definition

* fix up input-output plumbing

* fix tests

* Update agenthub/micro/math_agent/agent.yaml

Co-authored-by: Boxuan Li <liboxuan@connect.hku.hk>

* Update agenthub/delegator_agent/prompt.py

Co-authored-by: Boxuan Li <liboxuan@connect.hku.hk>

* Update agenthub/delegator_agent/prompt.py

Co-authored-by: Boxuan Li <liboxuan@connect.hku.hk>

* remove prompt.py

* fix lint

* Update agenthub/micro/postgres_agent/agent.yaml

Co-authored-by: Boxuan Li <liboxuan@connect.hku.hk>

* Update agenthub/micro/postgres_agent/agent.yaml

Co-authored-by: Boxuan Li <liboxuan@connect.hku.hk>

* fix error

---------

Co-authored-by: Boxuan Li <liboxuan@connect.hku.hk>
Co-authored-by: Engel Nyst <enyst@users.noreply.github.com>
2024-04-24 17:46:14 -04:00
..
agent
Microagents and Delegation (#1238)
2024-04-24 17:46:14 -04:00
auth
feat: improve credential validation and simplify response in get_token (#1155)
2024-04-18 07:00:24 -04:00
session
feat: support controlling agent task state. (#1094)
2024-04-18 11:09:00 +00:00
__init__.py
feat: websocket connection management and sandbox bound to session. (#559)
2024-04-05 12:19:52 -05:00
listen.py
feat: support tls. #1234 (#1248)
2024-04-21 14:50:48 -04:00
README.md
Add Docker DOOD setup (#1023)
2024-04-15 14:19:02 +02:00

README.md

OpenDevin Server

This is a WebSocket server that executes tasks using an agent.

Install

Follow the instructions in the base README.md to install dependencies and set up.

Start the Server

uvicorn opendevin.server.listen:app --reload --port 3000

Test the Server

You can use websocat to test the server: https://github.com/vi/websocat

websocat ws://127.0.0.1:3000/ws
{"action": "start", "args": {"task": "write a bash script that prints hello"}}

Supported Environment Variables

LLM_API_KEY=sk-... # Your OpenAI API Key
LLM_MODEL=gpt-3.5-turbo-1106 # Default model for the agent to use
WORKSPACE_BASE=/path/to/your/workspace # Default path to model's workspace

API Schema

There are two types of messages that can be sent to, or received from, the server:

  • Actions
  • Observations

Actions

An action has three parts:

  • action: The action to be taken
  • args: The arguments for the action
  • message: A friendly message that can be put in the chat log

There are several kinds of actions. Their arguments are listed below. This list may grow over time.

  • initialize - initializes the agent. Only sent by client.
    • model - the name of the model to use
    • directory - the path to the workspace
    • agent_cls - the class of the agent to use
  • start - starts a new development task. Only sent by the client.
    • task - the task to start
  • read - reads the content of a file.
    • path - the path of the file to read
  • write - writes the content to a file.
    • path - the path of the file to write
    • content - the content to write to the file
  • run - runs a command.
    • command - the command to run
    • background - if true, run the command in the background
  • kill - kills a background command
    • id - the ID of the background command to kill
  • browse - opens a web page.
    • url - the URL to open
  • recall - searches long-term memory
    • query - the query to search for
  • think - Allows the agent to make a plan, set a goal, or record thoughts
    • thought - the thought to record
  • finish - agent signals that the task is completed

Observations

An observation has four parts:

  • observation: The observation type
  • content: A string representing the observed data
  • extras: additional structured data
  • message: A friendly message that can be put in the chat log

There are several kinds of observations. Their extras are listed below. This list may grow over time.

  • read - the content of a file
    • path - the path of the file read
  • browse - the HTML content of a url
    • url - the URL opened
  • run - the output of a command
    • command - the command run
    • exit_code - the exit code of the command
  • recall - the result of a search
    • query - the query searched for
  • chat - a message from the user